Memo: On the Matter of Whose Balance Sheet This Actually Is

TO: Capital Allocation Committee
RE: Q3 planning assumptions, AI infrastructure line item
CLASSIFICATION: Internal, do not circulate outside finance

We need to talk about the $750 billion.

That's the number the four hyperscalers have now committed to AI capital expenditure for calendar 2026, up from the figures we modeled at the start of the year. Moody's has it crossing $1 trillion in 2027. McKinsey, ever the optimist's optimist, has global AI-data-center capex reaching $7 trillion by 2030. I want to flag, calmly and for the record, that none of these numbers were generated by asking whether the revenue exists to justify them. They were generated by asking how much compute the roadmap requires and then finding the financing.

That distinction matters more than anyone in this building currently wants to admit.

Here is the part of the model I keep returning to. S&P 500 consensus earnings growth for 2026 sits at 24%, per the sell-side aggregate — a number analysts themselves are calling, and I'm quoting the framing here, "significant but likely achievable." Fine. Now look at what happens after: 16% for 2027, 12.5% for 2028. That's not deceleration, that's a cliff dressed up in a business-casual font. The market is being asked to underwrite a trillion-dollar-a-year infrastructure buildout on the assumption that earnings growth halves twice in a row, and equities are pricing none of that compression in. The Dow closed above 53,000 for the first time this month. That's not a market discounting a deceleration curve. That's a market extrapolating the numerator and ignoring the denominator entirely.

Committee should also note: this capex is increasingly not being funded out of free cash flow. It is being funded, in growing part, by debt, by off-balance-sheet vehicles, and by circular financing arrangements between chipmakers and their own customers that read less like a supply chain and more like a hall of mirrors. Micron and SanDisk both posted outsized single-session pops this month — 4.5% and 7.6% respectively — on the same day oil fell and Iran tensions flared, which tells you the semiconductor complex has become largely decoupled from the macro tape it theoretically trades inside of. When a sector stops reacting to war and starts only reacting to memory bottleneck headlines, that sector has become a closed system. Closed systems don't correct gradually. They correct all at once, when the financing assumption underneath them changes.

Which brings me to rates. Developed-market yields are, per the desk's own outlook language, "near their highest levels of this cycle," with the expectation they slide lower "slowly." Slowly is doing a lot of work in that sentence. The entire AI capex thesis is a duration bet dressed as a growth story — it assumes cheap-enough financing stays available long enough for the revenue to catch up to the spend. Every basis point the 10-year holds above where the models assumed is a basis point subtracted from somebody's IRR on a data center that hasn't finished being poured yet.

I raise all this not to argue we exit the trade. We can't; every name on our book with AI-infrastructure exposure has been the alpha source of the last eighteen months, and Committee correctly overruled my caution in Q1 on that basis. I raise it because the structural risk here isn't a bad quarter from one hyperscaler. It's that four companies have effectively become the marginal buyer of the entire semiconductor and power-generation value chain simultaneously, and any one of them trimming guidance by even a routine, unremarkable amount now has second-order effects on utilities, industrial suppliers, and construction names that had nothing to do with AI six quarters ago and are now fully levered to its capex cycle continuing on schedule.

That's not diversification. That's concentration wearing a diversification costume, and it's sitting inside index products that half this firm's clients believe are "balanced."

Recommend the committee begin stress-testing the 2027 earnings assumption specifically — not the 2026 number, which is close enough to done that revising it does no good — against a scenario where hyperscaler capex growth merely flattens, not reverses. Flattens. I'm not asking anyone to model a bust. I'm asking what happens to the names currently priced for perpetual acceleration if the acceleration simply stops accelerating.

Nobody has run that scenario yet. Someone should, before the market runs it for us.